Adversarial Robustness of Self-Supervised Learning in Vision
Keywords: Self Supervised Learning, Adversarial Attacks, Robustness
TL;DR: We investigate adversarial robustness of the various SSL models in ImageNet, transfer learning, segmentation and detection.
Abstract: Self-supervised learning (SSL) has advanced significantly in visual representation learning, yet large-scale evaluations of its adversarial robustness remain limited. In this study, we evaluate the adversarial robustness of seven SSL models and one supervised model across a range of tasks, including ImageNet classification, transfer learning, segmentation, and detection. Our findings demonstrate that SSL models generally exhibit superior robustness to adversarial attacks compared to their supervised counterpart on ImageNet, with this advantage extending to transfer learning in classification tasks. However, this robustness is less pronounced in segmentation and detection tasks. We also explore the role of architectural choices in model robustness, observing that their impact varies depending on the SSL objective. Finally, we assess the effect of extended training durations on adversarial robustness, finding that longer training may offer slight improvements without compromising robustness. Our analysis highlights promising directions for enhancing the adversarial robustness of visual self-supervised representation systems in complex environments.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 12514
Loading